Phase diagram of congested traffic flow: an empirical study

Phase Diagram of Congested Traffic Flow: An Empirical Study

o :
o .
o ■

(N ;

3 ■
<■

>
00

o

O
O

i

c

o
o

13

H. Y. Lee 1 , H.-W. Lee 2 , and D. Kim 3
1 Department of Physics, Massachusetts Institute of Technology, Cambridge, MA 02139
2 School of Physics, Korea Institute for Advanced Study, 207-43 Cheongryangri-dong, Dongdaemun-gu, Seoul 130-012, Korea

3 School of Physics, Seoul National University, Seoul 151-742, Korea

We analyze traffic data from a highway section containing one effective on-ramp. Based on
two criteria, local velocity variation patterns and expansion (or nonexpansion) of congested regions,
three distinct congested traffic states are identified. These states appear at different levels of the
upstream flux and the on-ramp flux, thereby generating a phase digram of the congested traffic
flow. Observed traffic states are compared with recent theoretical analyses and both agreeing and
disagreeing features are found.

PACS numbers: 89.40. +k, 05.70.Fh, 45.70.Vn, 47.55.Kf

In the last decade, there has been growing interest in
the traffic flow, which is partly motivated by the fact that
the traffic flow is an easily perceivable realization of heav-
ily studied driven nonequilibrium systems |ffl). Another
important motivation is the hope that complex traffic
behaviors may be understood with the help of physical
approaches. Such understanding can be used to optimize
traffic and even to forecast traffic situations.

A reasonable first step along this line of thinking will
be the classification of distinct traffic states and separate
investigation of their properties. Various traffic models
are proposed || and compared with real traffic data || .
With the help of these models, free flow and so called
wide traffic jams are well understood. On the other
hand, nature of congested traffic flow (or synchronized
flow), which appears near road inhomogeneities mostly,
yet remains unclear despite various empirical ^|-|(| and
theoretical |f7Hll|] efforts.

A recent theoretical study || proposed an intriguing
possibility that the congested traffic flow may not be a
single dynamic phase but rather a collection of multiple
phases, each of which is realized under different condi-
tions. Similar conclusion is also reported from the inves-
tigation of another theoretical model JhJ. In the empir-
ical investigations, however, although qualitatively dis-
tinct congested traffic states are reported |@,|2|, no em-
pirical evidence is found for the existence of any char-
acteristic parameters that distinguish their appearance
conditions ||.

In this paper, we report empirical investigation of traf-
fic congestion in a highway section containing one effec-
tive on-ramp. Details of the section are given in our
preliminary report [ p^| , so we provide here only a brief
description (Fig. [I]) . All ramps are connected to the out-
ermost lane (lane 4) and a stretch of lane divider (from
x = 3.5 to 8.3 km, dashed line in Fig. [I]) blocks lane
change from the two outer lanes (lane 3 and 4) to the two
inner ones (lane I and 2) and vice versa. In a short road
segment near the end of the lane divider (from x = 8.3
km to the location of the detector D9 approximately),

many vehicles in the outer lanes switch into the inner
lanes, which is also enhanced by vehicle flux through the
on-ramp ON3 at x = 8.6 km. As a result, this segment
works as an effective on-ramp region for the traffic flow
of the inner lanes and traffic congestion often occurs in
the road section with the lane divider, where the inner
lanes are decoupled from the outer ones.

We investigate the traffic congestion in the inner lanes
using the 30 second averaged traffic data from June to
September, 1999 (a total of 107 days, a much larger data
set compared to 14 days in Ref. All quantities

below are averaged over the two inner lanes. For each
realization of a particular congested traffic state, which
is stably maintained about 30 min or longer, the effec-
tive ramp flux / rmp [defined as the difference between
two flux values measured at the detectors DIO and D7,
q(D10) — <z(D7)] and the upstream flux / up [q(Dn) is used
when the congestion extends up to Dn + 1] are averaged
over the time interval of its duration and the resulting av-
erage values (/ rmp ), (/up) are marked in Fig. ||(a). Note
that three congested traffic states, which we call CT2,
CT4, and CT5 states, respectively (see below), occupy
distinct regions, providing a supporting evidence for the
prediction [p||Ic|] that / up and / rmp are characteristic pa-
rameters of congested traffic flow. Thus Fig. ^(a) be-
comes an empirical phase diagram of the congested traffic
flow. Fig. |^(b) shows an alternative phase diagram ob-
tained from 10 min averaging of the flux values instead.
Two phase diagrams are qualitatively the same.

Three congested traffic states, CT2, CT4, and CT5,
are classified according to the two criteria given below.
It is previously reported j|] that the congested region
may consist of backward (towards upstream) traveling
clusters and the size of the clusters grow spontaneously
during their backward propagation. As a result, large
amplitude oscillation of velocity develops spontaneously.
On the other hand, recent theories |],[l(J predict that
large velocity oscillation may or may not develop. Thus
our criterion (i) is whether such spontaneous growth of
velocity oscillation appears (CT5) or not (CT2, CT4).

f

This criterion can be examined by comparing temporal
variation of velocity at different detectors.

In the same theoretical works, both expanding and
nonexpanding traffic states are predicted, and our crite-
rion (ii) is whether the congested region expands mono-
tonically (CT4, CT5) or not (CT2). Mathematically the
expansion rate of the congested region is proportional to
the degree of flux mismatch / up + / rmp — /down, where
/down measures the outflow from the congested region
[g(D10) is used]. Thus the comparison of (/ up ) + (/rm P )
and (/down) can be used as a more objective application
of the criterion (ii).

In all three states, the density-flow relations show fluc-
tuating behavior (such as Fig. 2(c) in Ref. [||) and the
velocity variations in the lane 1 and 2 are synchronized
(such as Fig. 2 in Ref. (5)). Since these properties are
already reported in other publications, we do not present
similar figures here. Below we discuss other properties of
the congested traffic states.

Fig. ||(a) depicts the density profile p(x,t) of the
CT2 state and its evolution with time, where p{x, t) is
evaluated by q(x,t) /u(x,t) and u(x,t) is the harmonic
mean velocity Jl4j over 30 s intervals. The sponta-
neous growth of the velocity oscillation mentioned in
the criterion (i) does not appear. The flux mismatch
(/imp) + (/up) - (/down) is negligible and the congested
region does not expand. Empirical data also show that
the length of the congested region increases with (/ rm p)
and its dependence on (/ up ) is rather weak |15|.

Fig. ||(b) shows the CT4 state. The spontaneous
growth of the velocity oscillation does not appear. Due to
the large flux mismatch (typical value of (/ up ) + (/rmp) —
(/down) is 500-600 veh/h), the congested region expands
with time. The expansion rate ranges from 3 to 9 km/h
and increases with increasing flux mismatch. An inter-
esting property is that the outflow (/down) is practically
independent of (/ up ) and (/ U p), and remains almost uni-
versal. The average of (/down) over the 28 realizations of
the CT4 state in Fig. |(a) is 2010 veh/h and its standard
deviation is about 65 veh/h, which is much smaller than
the spread of (/ up ) and (/rmp)- To our knowledge, this
is the first empirical indication of the universal outflow
from the congested flow near an on-ramp | jl6|[ .

Fig. ||(c) portrays the CT5 state. An important fea-
ture of the CT5 state is the spontaneous growth of the
velocity oscillation inside the congested region. Fig. |]
shows the temporal variation of the velocity at D5 and
D6. The graph for D6 is shifted to the right by 5 min
for comparison. Note that the velocity evolutions at the
two detectors are correlated after the time shift for D6,
and that the velocity oscillates with larger amplitude at
D5. These features imply that velocity wave propagates
backward (towards upstream) and its amplitude grows
during its propagation.

Regarding the criterion (ii), the congested region ex-
pands as shown in Fig. ^|(c). The upstream front of the

congested region initially locates between D5 and D6 and
later between D4 and D5. We mention that the flux at
D4 remains quasi-stationary during the depicted time in-
terval. Thus the expansion is not due to the increase
of /up but due to the flux mismatch. Compared to the
CT4 state, however, it turns out that the flux mismatch
(typically 200-250 veh/h) is considerably smaller, which
implies a slower expansion of the congested region. We
estimate the expansion rate by the flux mismatch divided
by the density difference at the upstream front of the
congested region, and find it ranges 2-4 km/h 0. The
outflow of this state is not universal.

It is interesting to compare empirically identified states
with theoretically predicted states fi-|lOt|. Application
of the criteria (i,ii) to both empirical and theoretical
states leads to the following pairing between empirical
and theoretical states: the CT2 state with the theoreti-
cally predicted pinned localized cluster (PLC) state, the
CT4 state with the homogeneous congested traffic (HCT)
state, and the CT5 state with the oscillating congested
traffic (OCT) state.

The pairing motivates further comparison between the
paired states. For the CT4-HCT pair, we note that the
universal outflow is predicted for the HCT state |)| and
the same property is observed for the CT4 state, which
strongly motivates the identification of the CT4 state
with the HCT states. For the CT5-OCT pair, on other
hand, while the outflow is universal for the OCT state |j) ,
it is not for the CT5 state. Thus properties of these two
states are only in partial agreement. And for the CT2-
PLC pair, we find one quantitative difference: the con-
gested region of the CT2 state is considerably wider than
that of the PLC state in Refs. §|l(]]. The resolution of
this discrepancy may require improved traffic theories.

We also compare the theoretical and empirical phase
diagrams. It is predicted in Ref. || that when the up-
stream flux and the ramp flux are maintained at strictly
constant values, / up and / rm p, respectively, the phase
boundary between the flux matching and flux mismatch-
ing states is practically identical to the stability bound-
ary of the free flow, below which the free flow can remain
linearly stable, and given by the line /u P + a/ r mp = <3out,
where a = 1 and Qout is a constant whose value is almost
identical to the universal outflow of the HCT state (a is
predicted to be a little larger than 1 in Ref. [fio||).

Empirical determination of the free flow stability
boundary is not an easy task since the free flow near
the boundary is quite vulnerable to fluctuations. In
the empirical phase diagram [Fig. |](a)], the dashed line
((/up) +a(/rm P ) = Qout, where a rj 1.3 and Q ut ~ 2100
veh/h) is an empirical estimation of the stability bound-
ary. Here the values of a and Qout are reliable up to
their first significant digits and their second significant
digits are rather uncertain. Within this accuracy, a is
approximately one and the value of Qout is close to the

2

universal outflow of the CT4 state. It is also pleasing
to note that this line divides (except for a small num-
ber of data points) the flux matching and mismatching
states, in agreement with the theoretical prediction
This feature is robust and independent of details of the
boundary estimation method although the values of a
and Q out depend on the details. We also note that in a
later, more refined theory |l2|], it is predicted that the
PLC and OCT states overlap weakly in the phase dia-
gram. Recalling the pairing CT2-PLC and CT5-OCT,
the weak overlap of the CT2 and CT5 states in Fig. |^(a)
is in agreement with this prediction. Also the locations
of the empirical and theoretical overlap regions in the
phase diagram are similar.

Regarding the phase boundary between the two ex-
panding traffic states (CT4,CT5), while the recent the-
ory H predicts the line / rmp = const, as the phase bound-
ary between the HCT and OCT states, Fig. ||(a) shows
that the boundary between the CT4 and CT5 states is
better described by (/ up ) = const. Thus as for the phase
boundary between the expanding states, empirical results
and the theoretical prediction do not agree.

We next discuss an implicit but important conceptual
implication of the recent theories where traffic

phases are identified with resulting final states that traf-
fic flow dynamics lead to after sufficient transient time. In
the language of nonlinear dynamics, all these states cor-
respond to stable attractors |l8| of traffic flow dynamics.
An implication of this concept is that for a given external
condition (such as / rm p, /up and ramp geometry), the re-
sulting traffic state is independent of details of the initial
traffic state or its "evolution history" Jl9| .

This idea can be tested empirically, for example, by
comparing realizations with the same external conditions
but with different evolution histories. Fig. compares the
(time averaged) velocity profile of two realizations of CT2
state, both with almost the same (/ rm p) and (/ U p)- But
their evolution histories are different: one has evolved
from the free flow and the other from the CT5 state (in-
set in Fig. ||). Note that two profiles almost overlap with
each other despite qualitatively different evolution his-
tories. This insensitivity to the evolution history con-
firms the conceptual implication of the recent theoretical
works.

Obviously there is a fundamental difference between
theoretical and empirical situations: while theories |?] |T(]]
assume strictly constant upstream and ramp fluxes, in
real traffic the fluxes fluctuate all the time. Thus in order
to make a rigorous comparison with theories, a good un-
derstanding of fluctuation effects is necessary. However
such an understanding is not available at present, so in
this paper, the fluctuation effects are essentially ignored
and the time averaged values (/ rm p) and (/ up ) are instead
used in the analysis. Hence in a strict sense, the present
investigation is only a correlation analysis between the
time averaged flux values and the traffic states that are

maintained for a sufficiently long time. In retrospect,
however, it turns out that many empirical results can be
explained by the theories if one accepts that (f lmp ) and
(/up) play the roles of the constant ramp flux and up-
stream flux, respectively, assumed in the theories. Hence
it appears that the neglect of the fluctuation effects can
be justified at least a posteriori.

Its justification can be partly understood from two ar-
guments: firstly, while the fluctuation effects are usually
crucial near the phase boundary, traffic states far from
the phase boundary are relatively insensitive to fluctu-
ations. Thus when a traffic state occupies a sufficiently
wide region in the phase diagram, the use of the time
averaged flux values for the phase diagram can be jus-
tified. Secondly, dominant contributions to flux fluctu-
ations come from short time scale of order 1 min, and
in the time scale of about ten min or longer, flux varia-
tion is almost quasi-stationary. Thus provided that the
relevant time scale of a traffic state (not very close to
the phase boundary) is longer than I min, the short time
scale fluctuations will be effectively averaged out by the
traffic dynamics itself and can be neglected indeed.

In support of these arguments, we find that our results
are not sensitive to the length of the averaging time inter-
val as long as it is sufficiently longer than 1 min. When
the same analysis is repeated with 10 min averaging,
Fig. |^(b) is obtained. Note that this new phase diagram
is qualitatively the same as the former one [Fig. ||(a)] and
all discussions above remain unchanged. Changes occur
only in a quantitative level. With the 10 min averaging,
the standard deviation of the outflow (/down) from the
CT4 state increases to about 85 veh/h.

We remark on a few details of the analysis. Firstly,
there exists another on-ramp (ON4) at 2.3 km down-
stream from ON3. Sometimes vehicle flux through this
on-ramp causes traffic congestion and the resulting con-
gestion extends to the region studied in this paper. All
such events are excluded from the present analysis to fo-
cus on one particular inhomogeneity po|| . Secondly, in
the analyzed road section, there are two spots (one be-
tween D5 and D6 and the other near D7), where the
lane divider is imperfect. All time intervals with non-
negligible vehicle fluxes through those spots are not in-
cluded in the analysis.

In summary, three congested traffic states are identi-
fied based on local velocity variation patterns and ex-
pansion (or nonexpansion) of the congested region. It
is found that the appearance of these congested traffic
states is strongly correlated with the time averaged flux
values, (/up) and (/ rmp ), providing a strong supporting
evidence to the prediction ||[l0| that these flux values
are characteristic parameters of the congested traffic flow.
An empirical phase diagram is constructed and compared
with theoretical predictions. The prediction on the phase
boundary between the flux matching and mismatching
states is consistent with the empirical phase diagram

3

and the prediction of the universal outflow are confirmed.
However some deviations from theoretical predictions are
also found. Lastly we mention that there exist regions in
the (/rmp)-(/up) plane which are not probed. Thus it is
possible that additional congested traffic states exist in
those regions.

We thank Young-lhn Lee and Seung Jin Lee for pro-
viding the traffic data, Sung Yong Park for fruitful dis-
cussions. This work was supported in part by the BK2f
Project of Korean Ministry of Education.

[1]
[2]

[3]

[4]

[5]

[6]
[7]

[8]
[9]
[10]

[11]
[121

m

[15]

B. Shmittmann and R. K. Zia, in Phase Transitions and
Critical Phenomena, Vol. 17, edited by C. Domb and J.
Lebowitz (Academic, New York, 1995).
K. Nagel and M. Schreckenberg, J. Phys. I (France) 2,
2221 (1992); B. S. Kerner and P Konhauser, Phys. Rev.
E 48, R2335 (1993); M. Bando, K. Hasebe, A. Nakayama,

A. Shibata, and Y. Sugiyama, ibid. 51, 1035 (1995); S.
Krauss, P. Wagner, and C. Gawron, ibid. 55, 5597 (1997).

B. S. Kerner and H. Rehborn, Phys. Rev. E 53, R1297
(1996), and references cited therein.

B. S. Kerner and H. Rehborn, Phys. Rev. E 53, R4275

(1996) .

B. S. Kerner and H. Rehborn, Phys. Rev. Lett. 79, 4030

(1997) .

B. S. Kerner, Phys. Rev. Lett. 81, 3797 (1998).

H. Y. Lee, H.-W. Lee, and D. Kim, Phys. Rev. Lett. 81,

1130 (1998).

D. Helbing and M. Treiber, Phys. Rev. Lett. 81, 3042

(1998) .

D. Helbing, A. Hennecke, and M. Treiber, Phys. Rev.
Lett. 82, 4360 (1999).

H. Y. Lee, H.-W. Lee, and D. Kim, Phys. Rev. E 59,
5101 (1999).

E. Tomer, L. Safonov, and S. Havlin, Phys. Rev. Lett.
84, 382 (2000).

M. Treiber, A. Hennecke, and D. Helbing, cond

mat/0002177

H. Y. Lee, H .-W. Lee, and D. Kim, unpublished ( |cond-

mat/9905292| )

The harmonic mean velocity u [l/u = (1/N) ^ =1 l/vj]

is used instead of v [v = (1/N) Vj] since q/v sys-

tematically underestimates the density when the veloc-
ity fluctuations are significant (see for example L. Neu-
bert, L. Santen, A . Schadschneider, M. Schreckenberg,
cond- mat /9905216] ) whereas q/u takes a better account
of the fluctuation effects. Also it can be shown that when
both density and flux are temporally averaged [for ex-
(1/T)/_ T/2

ample, p(x, t)

-T/2

dt'V.<5(x 3 (* + t')

q(x,t) = (1/T) S_' T/2 df J2j «j(t + t')8(xj(t + O - *)],
the ratio q/p is exactly equal to u.

Usually the boundary between the congested region and
the upstream free region locates between detectors. But

due to the dependence of the congested region size on
(/rmp), the boundary will locate at a particular detector
when (/rmp) is close to a certain special value. In such a
case, small fluctuations of the boundary can cause large
amplitude fluctuations of velocity at that special detec-
tor since the velocity varies greatly across the bound-
ary. Such large velocity fluctuations are indeed observed
in narrow ranges of (/rm P ) (for example at D6 when
(/rmp) = 330 ± 30 veh/h). We remark that these fluc-
tuations should be distinguished from the velocity oscil-
lations [^| in the criterion (i). While the former originates
from the boundary fluctuations and is localized near the
boundary, the latter grows continuously inside the con-
gested region and is extended over relatively wide con-
gested region.

[16] At 30 s scale, /down [g(D10)] fluctuates within the range
2010 ± 140 veh/h, which is equivalent to 16.8 ± 1.2
veh/30 s. Here the inaccuracy of ±1.2 veh/30 s is quite a
small number considering that the vehicle number count-
ing by detectors always results in integer numbers and
thus the inaccuracy of ~ ±1 veh/30 s is unavoidable in
30 s data.

[17] It may be hard to detect slow expansion from the 3d
density plot due to the finite (~ 1 km) spatial resolution
given by the average detector spacing. For this reason,
this state is incorrectly interpreted as a nonexpanding
state in our preliminary report |l^] (using only 14 day
traffic data), where a different name, CT1, is used for
the same state.

[18] E. A. Jackson, Perspectives of Nonlinear Dynamics
(Cambridge University Press, Cambridge, England,
1989), Vol. 1.

[19] When there exist multiple (locally) stable attractors for
a given external condition, initial traffic state is not com-
pletely irrelevant. But even in this case, the initial state
affects only the choice of a particular attractor as a final
state. Or in the language of nonlinear dynamics, details
of the initial state are irrelevant as long as the initial
state remains within the basin of attraction jig] .

[20] It is found that the on-ramp flux through ON4 may gen-
erate still another type of congested traffic state (called
CT3 state in Ref. ) , which is not included in the
present analysis since this state appears under different
road geometry.

Dl D2 D3 D4

D5 D6 D7 D8D9D10D11

0.5 1.8 3.1 3.7

5.8 6.9 7.6 8.5 9.2 9.7 10.9 X (km)

El ONI

E2 ON2E3 0N3 E4 ON4

1.7 3.2 6.1 7.4 8.0 8.6 9.5 10.9 X fkm)

Lee, Lee, Kim Fig.l

4

FIG. 1. Schematic diagram of a road section in the
Olympic Highway in Seoul. Locations of detectors (Dn),
on-ramps (ONn), and off-ramps (En) are marked. The dashed
line in the middle denotes the lane divider and the arrow in-
dicates the driving direction.

2000

5 1600
>

CT2 *
CT4 I

CT5 v
FREE -

* *

1200

800

(a)

200

400 600

<f >(veh/h)

800

1000

Lee, Lee, Kim Fig. 2(a)

2000

CT2 X
CT4 I

CT5 v

■ t:

1000

(veh/h)

Lee, Lee, Kim Fig. 2(b)

FIG. 2. (a) Empirical phase diagram of the congested
traffic flow, (/up) and (/, mp ) represent the average upstream
and on-ramp flux values over the time interval during which
a particular congested traffic state is maintained (also lane
averaged). The dashed line is an empirical estimation of the
free flow phase boundary below which the free flow can remain
linearly stable, (b) Same diagram using 10 min averaged flux
values.

p (veh/km)

14:40 4

x (km) (a)

Lee, Lee, Kim Fig. 3(a)

p (veh/km)

7:15^4

x(km) (b)

Lee, Lee, Kim Fig. 3(b)

p (veh/km)

16:45

X (km) (c)

Lee, Lee, Kim Fig. 3(c)

FIG. 3. The 3d density profile of the CT2 state (a), CT4
state (b), and CT5 state (c).

5

13:35

13:40

13:45

13:50

Lee, Lee, Kim Fig. 4

FIG. 4. Spontaneous growth of the velocity oscillation
inside the congested region of the CT5 state. The velocity
evolution at D6 is shifted to the right by 5 min for clear com-
parison.

A

>

v

CT2
CT5
FREE

300 600 900

<f rmp > (veh/h)

x (km)
Lee, Lee, Kim Fig. 5

FIG. 5. Time averaged velocity profiles of two realizations
of the CT2 state with almost identical (/ up ) and {/ rmp ) but
with different evolution histories (solid line for the realization
1 and dashed line for the realization 2). Inset: While the
realization 1 has developed from the CT5 state via a phase
transition, the realization 2 has evolved from the free flow.

G